cancel
Showing results for 
Search instead for 
Did you mean: 
cancel
Showing results for 
Search instead for 
Did you mean: 

Prescriptive analytics

Re: Prescriptive analytics

I see. The link seems helpful so i will most probably try it.

 

I am planing to work with time series datasets so the question is:

Does thingworx prescriptive thing work with time series datasets?

 

I was reading up on how prescriptive analytics can in real time suggesting actions to benefit from the predictions and show the implications of each decision option on the end goal eg. profit so i wonder if thingworx has such functionality.

 

 

 

Re: Prescriptive analytics

I haven't tried this before. Technically, time series are transformed and handled like non-time series data, because the learner algorithms don't really support the concept of "time". Question is, how much time do you need to know in advance to react, because this also defines the lookahead. I don't think it's possible to use prescription in a meaningful way with a lookahead>1 timeunits.

Re: Prescriptive analytics

In thingworx learners, does it support support rnn like gru and lstm as these seems to perform better in time series.

Re: Prescriptive analytics

No. There are a lot of specific ML algorithms around that perform better in certain situations, but that's not the goal of Thingworx Analytics. TWA addresses users that are NOT data scientists, but still want to get the job done. So PTC provides a couple of algorithms that perform best for the usual usecases the customers have and that don't require knowledge about machine learning. A data scientist-created model might out-perform a TWA-auto-created model, but at the cost of time and knowledge needed to create that model. If you are a data scientist or need to use learners other than the ones TWA has included or need a more elaborated model, you can always bring your own an import it to TWA, as long as it is in format PMML-4.3.

Re: Prescriptive analytics

Interestingly, i realize that you can get the the following error for wrong range input. How does the TWA determine the correct boundary for this dataset?

 

Failed to score: Invalid range specified for the continuous field [Cement].
Acceptable bound(s): [102.0,540.0]

Re: Prescriptive analytics

It's most probably the min and max value of the training data. The system can't make predictions beyond the range it has seen in the training.

Re: Prescriptive analytics

@Rocko 

Hi,

is it possible to briefly explain how the prescriptive analytics finds the values to maximise or minimise the goal? I got a little confused as the model i trained on is a mashup of neural network and some other algorithm. From what i understand, it acts like a black box with a lot of functions in the background so it seem to be impossible to find the global max or min.

Does Thingworx(prescriptive analytics) use brute force on all the possible values of the chosen levers to find the most optimal value?

Re: Prescriptive analytics

How it _exactly_ works is beyond my knowledge, but you can imagine it as looking for maxima on an (n-1) dimensional hyperplane by walking around on the surface. Each variable you define as a lever is one direction of walking around and looking for "better" outcomes.
Not sure it is looking for the global maximum, also it is not using the full vectorspace (brute-force) to find those values - this would be computationally too expensive, even if it was a discrete space. It's it using some smart way to find where it is best to look for better scores.
But the general approach is to repeatedly adjust the levers and see how the adjusted input scores.

Re: Prescriptive analytics

During pre-sales I often get asked by customers about the details on how TWA works under-the-hood. Is it documented somewhere, besides the info in help center?

I mean, I can explain DS concepts to people, but the people asking such questions already have some DS expertise and they want to know what approaches are implemented in TW Analytics to try to estimate quality of the models and predictions they'll be getting, otherwise they're reluctant to trust the "black box" software.

Even if we persuate the decision maker and win the deal, having a sceptical DS guy who will be implementing the solution will not be benefical to the project.

Yes, predictions themselves is just one piece of TW and TWA functionality and the product has other major benefits which I communicate during pre-sales, but I couldn't find a good way to handle such kind of questions.